An official website of the United States government
Current deep-learning techniques for processing sets are limited to a fixed cardinality, causing a steep increase in computational complexity when the set is large. To address this, we have taken techniques used to model long-term dependencies from natural language processing and combined them with the permutation equivariant architecture, Set Transformer (STr). The result is Set Transformer XL (STrXL), a novel deep learning model capable of extending to sets of arbitrary cardinality given fixed computing resources. STrXL's extension capability lies in its recurrent architecture. Rather than processing the entire set at once, STrXL processes only a portion of the set at a time and uses a memory mechanism to provide additional input from the past. STrXL is particularly applicable to processing sets of high-throughput sequencing (HTS) samples of DNA sequences as their set sizes can range into hundreds of thousands. When tasked with classifying HTS prairie soil samples and MNIST digits, results show that STrXL exhibits an expected memory size-accuracy trade-off that scales proportionally with the complexity of downstream tasks, but, unlike STr, is capable of generalizing to sets of arbitrary cardinality.
more »
« less
This content will become publicly available on May 14, 2026
DNAGAST: Generative Adversarial Set Transformers for High-throughput Sequencing
High-throughput sequencing (HTS) is a modern DNA sequencing technology used to rapidly read thousands of genomic fragments from microorganisms given a sample. The large amount of data produced by this process makes deep learning, whose performance often scales with dataset size, a suitable fit for processing HTS samples. While deep learning models have utilized sets of DNA sequences to make informed predictions, to our knowledge, there are no models in the current literature capable of generating synthetic HTS samples, a tool which could enable experimenters to predict HTS samples given some environmental parameters. Furthermore, the unordered nature of HTS samples poses a challenge to nearly all deep learning architectures because they have an inherent dependence on input order. To address this gap in the literature, we introduce DNA Generative Adversarial Set Transformer (DNAGAST), the first model capable of generating synthetic HTS samples.We qualitatively and quantitatively demonstrate DNAGAST’s ability to produce realistic synthetic samples and explore various methods to mitigate mode-collapse. Additionally, we propose novel quantitative diversity metrics to measure the effects of mode-collapse for unstructured set-based data.
more »
« less
- Award ID(s):
- 1933925
- PAR ID:
- 10623700
- Publisher / Repository:
- University of Florida Press
- Date Published:
- Journal Name:
- The International FLAIRS Conference Proceedings
- Volume:
- 38
- ISSN:
- 2334-0754
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Current deep-learning techniques for processing sets are limited to a fixed cardinality, causing a steep increase in computational complexity when the set is large. To address this, we have taken techniques used to model long-term dependencies from natural language processing and combined them with the permutation equivariant architecture, Set Transformer (STr). The result is Set Transformer XL (STrXL), a novel deep learning model capable of extending to sets of arbitrary cardinality given fixed computing resources. STrXL’s extension capability lies in its recurrent architecture. Rather than processing the entire set at once, STrXL processes only a portion of the set at a time and uses a memory mechanism to provide additional input from the past. STrXL is particularly applicable to processing sets of highthroughput sequencing (HTS) samples of DNA sequences as their set sizes can range into hundreds of thousands. When tasked with classifying HTS prairie soil samples and MNIST digits, results show that STrXL exhibits an expected memory size-accuracy trade-off that scales proportionally with the complexity of downstream tasks, but, unlike STr, is capable of generalizing to sets of arbitrary cardinality.more » « less
-
Abstract MotivationHigh-throughput sequencing (HTS) is a modern sequencing technology used to profile microbiomes by sequencing thousands of short genomic fragments from the microorganisms within a given sample. This technology presents a unique opportunity for artificial intelligence to comprehend the underlying functional relationships of microbial communities. However, due to the unstructured nature of HTS data, nearly all computational models are limited to processing DNA sequences individually. This limitation causes them to miss out on key interactions between microorganisms, significantly hindering our understanding of how these interactions influence the microbial communities as a whole. Furthermore, most computational methods rely on post-processing of samples which could inadvertently introduce unintentional protocol-specific bias. ResultsAddressing these concerns, we present SetBERT, a robust pre-training methodology for creating generalized deep learning models for processing HTS data to produce contextualized embeddings and be fine-tuned for downstream tasks with explainable predictions. By leveraging sequence interactions, we show that SetBERT significantly outperforms other models in taxonomic classification with genus-level classification accuracy of 95%. Furthermore, we demonstrate that SetBERT is able to accurately explain its predictions autonomously by confirming the biological-relevance of taxa identified by the model. Availability and implementationAll source code is available at https://github.com/DLii-Research/setbert. SetBERT may be used through the q2-deepdna QIIME 2 plugin whose source code is available at https://github.com/DLii-Research/q2-deepdna.more » « less
-
Generative adversarial networks (GANs) are a technique for learning generative models of complex data distributions from samples. Despite remarkable advances in generating realistic images, a major shortcoming of GANs is the fact that they tend to produce samples with little diversity, even when trained on diverse datasets. This phenomenon, known as mode collapse, has been the focus of much recent work. We study a principled approach to handling mode collapse, which we call packing. The main idea is to modify the discriminator to make decisions based on multiple samples from the same class, either real or artificially generated. We draw analysis tools from binary hypothesis testing—in particular the seminal result of Blackwell [4]—to prove a fundamental connection between packing and mode collapse. We show that packing naturally penalizes generators with mode collapse, thereby favoring generator distributions with less mode collapse during the training process. Numerical experiments on benchmark datasets suggest that packing provides significant improvements.more » « less
-
Generative adversarial networks (GANs) are innovative techniques for learning generative models of complex data distributions from samples. Despite remarkable recent improvements in generating realistic images, one of their major shortcomings is the fact that in practice, they tend to produce samples with little diversity, even when trained on diverse datasets. This phenomenon, known as mode collapse, has been the main focus of several recent advances in GANs. Yet there is little understanding of why mode collapse happens and why recently proposed approaches are able to mitigate mode collapse. We propose a principled approach to handling mode collapse, which we call packing. The main idea is to modify the discriminator to make decisions based on multiple samples from the same class, either real or artificially generated. We borrow analysis tools from binary hypothesis testing—in particular the seminal result of Blackwell [6]—to prove a fundamental connection between packing and mode collapse. We show that packing naturally penalizes generators with mode collapse, thereby favoring generator distributions with less mode collapse during the training process. Numerical experiments on benchmark datasets suggests that packing provides significant improvements in practice as well.more » « less
